| 6 | |
| 7 | |
| 8 | class OPTModel(LLM): |
| 9 | model_name: str = "" |
| 10 | tokenizer: AutoTokenizer = None |
| 11 | model: OPTForCausalLM = None |
| 12 | |
| 13 | def __init__(self, huggingface_model_name: str) -> None: |
| 14 | super().__init__() |
| 15 | self.model_name = huggingface_model_name |
| 16 | self.tokenizer = AutoTokenizer.from_pretrained(self.model_name) |
| 17 | self.model = OPTForCausalLM.from_pretrained(self.model_name) |
| 18 | |
| 19 | @property |
| 20 | def _llm_type(self) -> str: |
| 21 | return self.model_name |
| 22 | |
| 23 | def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str: |
| 24 | |
| 25 | inputs = self.tokenizer( |
| 26 | prompt, |
| 27 | padding=True, |
| 28 | max_length=512, # 512 by default，tokenizer.model_max_length=1000000000000000019884624838656 |
| 29 | truncation=True, |
| 30 | return_tensors="pt" |
| 31 | ) |
| 32 | |
| 33 | inputs_len = inputs["input_ids"].shape[1] |
| 34 | |
| 35 | generated_outputs = self.model.generate( |
| 36 | inputs['input_ids'], |
| 37 | max_new_tokens=512, |
| 38 | ) |
| 39 | decoded_output = self.tokenizer.batch_decode( |
| 40 | generated_outputs[..., inputs_len:], skip_special_tokens=True, clean_up_tokenization_spaces=False) |
| 41 | |
| 42 | output = decoded_output[0] |
| 43 | return output |
| 44 | |
| 45 | @property |
| 46 | def _identifying_params(self) -> Mapping[str, Any]: |
| 47 | """Get the identifying parameters.""" |
| 48 | return {"model_name": self.model_name} |
| 49 | |
| 50 | if __name__ == "__main__": |
| 51 | llm = OPTModel("facebook/opt-350m") |